Abstract
MOTIVATION: Predicting drug–disease associations (DDAs) is essential for efficient drug repurposing. Although graph convolutional networks (GCNs) on heterogeneous drug–disease graphs are state-of-the-art, they often underutilize the rich, multi-modal data available for drugs, such as targets, enzymes, pathways, and chemical substructures. RESULTS: To address this, we introduce Multi-DDA, a novel framework that systematically integrates these multi-modal drug features into a dedicated learning branch. These enriched drug descriptors are hierarchically combined with the outputs of each graph convolution layer, allowing subsequent layers to selectively refine the most informative node representations. This multi-modal fusion creates more comprehensive drug and disease embeddings. The representations are then processed by a graph attention layer to weigh the importance of different node connections before a final Multi-Layer Perceptron predicts the association matrix. Evaluated on a benchmark dataset of 269 drugs and 598 diseases, Multi-DDA outperforms seven existing methods across key metrics—Area Under the Precision-Recall Curve (AUPR), Area Under the Receiver Operating Characteristic Curve (AUC), and Recall. The significant gains in AUPR and Recall demonstrate its enhanced capability to identify potential DDAs, offering a powerful tool for advancing personalized medicine and drug discovery. AVAILABILITY AND IMPLEMENTATION: The source code for Multi-DDA is freely available at https://github.com/dehghan1401/Multi-DDA